A Novel Polarimetric SAR Classification Method Integrating Pixel-Based and Patch-Based Classification

Abstract

A new polarimetric synthetic aperture radar (SAR) images classification method based on residual network (ResNet) and deep autoencoder (DAE) is proposed in this letter. The patch-based classification and pixel-based classification are well integrated to achieve better classification accuracy and clearer contour features. The patch-based classification results with ResNet and pixel-based classification results with DAE are obtained respectively. According to the results, a hybrid method combining the patch-based and the pixel-based classification is developed to determine the category label of each pixel. The attractive feature of the combined method is to take full use of the polarization scattering characteristics in each pixel and spatial information of the polarimetric SAR data. To verify the proposed method, SAR images from Chinese GaoFen 3 (GF-3) space-borne SAR systems are used and experiments are performed, which shows the proposed method can achieve high accuracy and maintain contour features simultaneously. Compared with existing classification methods, the new method has a better performance in classification accuracy and false alarm probability (FAP).